Scholarlink

In this paper, we introduce a BERT-based attention model integrated with a bidirectional LSTM (BiLSTM) for aspect-based sentiment analysis (ABSA) in the agricultural domain. ABSA is a granular sentiment analysis approach that identifies sentiment polarities specific to aspects or features within a text, providing more detailed insights than traditional sentiment analysis. The proposed model leverages the pre-trained BERT model to generate contextual embeddings, which are then enhanced by a multi-head attention mechanism to capture the relations between aspect and their surrounding context. A BiLSTM layer is subsequently used to refine these representations, aiding in the sequential understanding of sentiment. We evaluated our model on our specific datasets, which are publicly available on GitHub: an Amazon review dataset related to pesticide products, where it achieved an accuracy of 96.43%, and a Twitter dataset focused on pest and crop disease discussions, achieving an accuracy of 92.74%. Additionally, we tested our model on popular ABSA benchmark datasets: SemEval2014-Task4, SemEval2015-Task12, and SemEval2016-Task5 restaurants datasets, where it achieved accuracies of 90.30%, 94.90%, and 95.81%, respectively. These results demonstrate our model’s effectiveness across both domain-specific and general datasets, emphasizing its adaptability in ABSA tasks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A BERT-Based Attention Model with BiLSTM for Aspect-Based Sentiment Analysis in Agricultural Datasets

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ameni Chamekh,
  • Mariem Mahfoudh,
  • Khouloud Boukadi,
  • Walid Hamada

摘要

In this paper, we introduce a BERT-based attention model integrated with a bidirectional LSTM (BiLSTM) for aspect-based sentiment analysis (ABSA) in the agricultural domain. ABSA is a granular sentiment analysis approach that identifies sentiment polarities specific to aspects or features within a text, providing more detailed insights than traditional sentiment analysis. The proposed model leverages the pre-trained BERT model to generate contextual embeddings, which are then enhanced by a multi-head attention mechanism to capture the relations between aspect and their surrounding context. A BiLSTM layer is subsequently used to refine these representations, aiding in the sequential understanding of sentiment. We evaluated our model on our specific datasets, which are publicly available on GitHub: an Amazon review dataset related to pesticide products, where it achieved an accuracy of 96.43%, and a Twitter dataset focused on pest and crop disease discussions, achieving an accuracy of 92.74%. Additionally, we tested our model on popular ABSA benchmark datasets: SemEval2014-Task4, SemEval2015-Task12, and SemEval2016-Task5 restaurants datasets, where it achieved accuracies of 90.30%, 94.90%, and 95.81%, respectively. These results demonstrate our model’s effectiveness across both domain-specific and general datasets, emphasizing its adaptability in ABSA tasks.